1. Purpose and scope
Architecture turns the targets and the chosen option into a specification that design, verification, software, package and system teams can all build from. For an accelerator, the most consequential part is often not the compute array but the memory hierarchy and the interface software sees: the instruction set or command format, the memory model, synchronisation, and how the host and other chips talk to the accelerator.
The stage ends at architecture freeze. After it, changes go through change control with their effect on performance, software and schedule stated. Freezing the software interface here is what lets compiler, runtime and driver work proceed in parallel with hardware design.
12 more sections for registered members.